The present invention relates in general to an improved method for and apparatus of a computer data processing system; and in particular, to an improved high performance multithreaded computer data processing system and method embodied in the hardware of the processor.
The fundamental structure of a modern computer includes peripheral devices to communicate information to and from the outside world; such peripheral devices may be keyboards, monitors, tape drives, communication lines coupled to a network, etc. Also included in the basic structure of the computer is the hardware necessary to receive, process, and deliver this information from and to the outside world, including busses, memory units, input/output (I/O) controllers, storage devices, and at least one central processing unit (CPU), etc. The CPU is the brain of the system. It executes the instructions which comprise a computer program and directs the operation of the other system components.
From the standpoint of the computer's hardware, most systems operate in fundamentally the same manner. Processors actually perform very simple operations quickly, such as arithmetic, logical comparisons, and movement of data from one location to another. Programs which direct a computer to perform massive numbers of these simple operations give the illusion that the computer is doing something sophisticated. What is perceived by the user as a new or improved capability of a computer system, however, may actually be the machine performing the same simple operations, but much faster. Therefore continuing improvements to computer systems require that these systems be made ever faster.
One measurement of the overall speed of a computer system, also called the throughput, is measured as the number of operations performed per unit of time. Conceptually, the simplest of all possible improvements to system speed is to increase the clock speeds of the various components, particularly the clock speed of the processor. So that if everything runs twice as fast but otherwise works in exactly the same manner, the system will perform a given task in half the time. Computer processors which were constructed from discrete components years ago performed significantly faster by shrinking the size and reducing the number of components; eventually the entire processor was packaged as an integrated circuit on a single chip. The reduced size made it possible to increase the clock speed of the processor, and accordingly increase system speed.
Despite the enormous improvement in speed obtained from integrated circuitry, the demand for ever faster computer systems still exists. Hardware designers have been able to obtain still further improvements in speed by greater integration, by further reducing the size of the circuits, and by other techniques. Designers, however, think that physical size reductions cannot continue indefinitely and there are limits to continually increasing processor clock speeds. Attention has therefore been directed to other approaches for further improvements in overall speed of the computer system.
Without changing the clock speed, it is still possible to improve system speed by using multiple processors. The modest cost of individual processors packaged on integrated circuit chips has made this practical. The use of slave processors considerably improves system speed by off-loading work from the CPU to the slave processor. For instance, slave processors routinely execute repetitive and single special purpose programs, such as input/output device communications and control. It is also possible for multiple CPUs to be placed in a single computer system, typically a host-based system which services multiple users simultaneously. Each of the different CPUs can separately execute a different task on behalf of a different user, thus increasing the overall speed of the system to execute multiple tasks simultaneously. It is much more difficult, however, to improve the speed at which a single task, such as an application program, executes. Coordinating the execution and delivery of results of various functions among multiple CPUs is a tricky business. For slave I/O processors this is not so difficult because the functions are pre-defined and limited but for multiple CPUs executing general purpose application programs it is much more difficult to coordinate functions because, in part, system designers do not know the details of the programs in advance. Most application programs follow a single path or flow of steps performed by the processor. While it is sometimes possible to break up this single path into multiple parallel paths, a universal application for doing so is still being researched. Generally, breaking a lengthy task into smaller tasks for parallel processing by multiple processors is done by a software engineer writing code on a case-by-case basis. This ad hoc approach is especially problematic for executing commercial transactions which are not necessarily repetitive or predictable.
Thus, while multiple processors improve overall system performance, there are still many reasons to improve the speed of the individual CPU. If the CPU clock speed is given, it is possible to further increase the speed of the CPU, i.e., the number of operations executed per second, by increasing the average number of operations executed per clock cycle. A common architecture for high performance, single-chip microprocessors is the reduced instruction set computer (RISC) architecture characterized by a small simplified set of frequently used instructions for rapid execution, those simple operations performed quickly as mentioned earlier. As semiconductor technology has advanced, the goal of RISC architecture has been to develop processors capable of executing one or more instructions on each clock cycle of the machine. Another approach to increase the average number of operations executed per clock cycle is to modify the hardware within the CPU. This throughput measure, clock cycles per instruction, is commonly used to characterize architectures for high performance processors. Instruction pipelining and cache memories are computer architectural features that have made this achievement possible. Pipeline instruction execution allows subsequent instructions to begin execution before previously issued instructions have finished. Cache memories store frequently used and other data nearer the processor and allow instruction execution to continue, in most cases, without waiting the full access time of a main memory. Some improvement has also been demonstrated with multiple execution units with look ahead hardware for finding instructions to execute in parallel.
The performance of a conventional RISC processor can be further increased in the superscalar computer and the Very Long Instruction Word (VLIW) computer, both of which execute more than one instruction in parallel per processor cycle. In these architectures, multiple functional or execution units are provided to run multiple pipelines in parallel. In a superscalar architecture, instructions may be completed in-order and out-of-order. In-order completion means no instruction can complete before all instructions dispatched ahead of it have been completed. Out-of-order completion means that an instruction is allowed to complete before all instructions ahead of it have been completed, as long as a predefined rules are satisfied.
For both in-order and out-of-order completion of instructions in superscalar systems, pipelines will stall under certain circumstances. An instruction that is dependent upon the results of a previously dispatched instruction that has not yet completed may cause the pipeline to stall. For instance, instructions dependent on a load/store instruction in which the necessary data is not in the cache, i.e., a cache miss, cannot be completed until the data becomes available in the cache. Maintaining the requisite data in the cache necessary for continued execution and to sustain a high hit ratio, i.e., the number of requests for data compared to the number of times the data was readily available in the cache, is not trivial especially for computations involving large data structures. A cache miss can cause the pipelines to stall for several cycles, and the total amount of memory latency will be severe if the data is not available most of the time. Although memory devices used for main memory are becoming faster, the speed gap between such memory chips and high-end processors is becoming increasingly larger. Accordingly, a significant amount of execution time in current high-end processor designs is spent waiting for resolution of cache misses and these memory access delays use an increasing proportion of processor execution time.
And yet another technique to improve the efficiency of hardware within the CPU is to divide a processing task into independently executable sequences of instructions called threads. This technique is related to breaking a larger task into smaller tasks for independent execution by different processors except here the threads are to be executed by the same processor. When a CPU then, for any of a number of reasons, cannot continue the processing or execution of one of these threads, the CPU switches to and executes another thread. The term "multithreading" as defined in the computer architecture community is not the same as the software use of the term which means one task subdivided into multiple related threads. In the architecture definition, the threads may be independent. Therefore "hardware multithreading" is often used to distinguish the two uses of the term. Within the context of the present invention, the term multithreading connotes hardware multithreading to tolerate memory latency.
Multithreading permits the processors' pipeline(s) to do useful work on different threads when a pipeline stall condition is detected for the current thread. Multithreading also permits processors implementing non-pipeline architectures to do useful work for a separate thread when a stall condition is detected for a current thread. There are two basic forms of multithreading. A traditional form is to keep N threads, or states, in the processor and interleave the threads on a cycle-by-cycle basis. This eliminates all pipeline dependencies because instructions in a single thread are separated. The other form of multithreading, and the one considered by the present invention, is to interleave the threads on some long-latency event.
Traditional forms of multithreading involves replicating the processor registers for each thread. For instance, for a processor implementing the architecture sold under the trade name PowerPC.TM. to perform multithreading, the processor must maintain N states to run N threads. Accordingly, the following are replicated N times: general purpose registers, floating point registers, condition registers, floating point status and control register, count register, link register, exception register, save/restore registers, and special purpose registers. Additionally, the special buffers, such as a segment lookaside buffer, can be replicated or each entry can be tagged with the thread number and, if not, must be flushed on every thread switch. Also, some branch prediction mechanisms, e.g., the correlation register and the return stack, should also be replicated. Fortunately, there is no need to replicate some of the larger functions of the processor such as: level one instruction cache (L1 I-cache), level one data cache (L1 D-cache), instruction buffer, store queue, instruction dispatcher, functional or execution units, pipelines, translation lookaside buffer (TLB), and branch history table. When one thread encounters a delay, the processor rapidly switches to another thread. The execution of this thread overlaps with the memory delay on the first thread.
Existing multithreading techniques describe switching threads on a cache miss or a memory reference. A primary example of this technique may be reviewed in "Sparcle: An Evolutionary Design for Large-Scale Multiprocessors," by Agarwal et al., IEEE Micro Volume 13, No. 3, pp. 48-60, June 1993. As applied in a RISC architecture, multiple register sets normally utilized to support function calls are modified to maintain multiple threads. Eight overlapping register windows are modified to become four non-overlapping register sets, wherein each register set is a reserve for trap and message handling. This system discloses a thread switch which occurs on each first level cache miss that results in a remote memory request. While this system represents an advance in the art, modern processor designs often utilize a multiple level cache or high speed memory which is attached to the processor. The processor system utilizes some well-known algorithm to decide what portion of its main memory store will be loaded within each level of cache and thus, each time a memory reference occurs which is not present within the first level of cache the processor must attempt to obtain that memory reference from a second or higher level of cache.
It should thus be apparent that a need exists for an improved data processing system which can reduce delays due to memory latency in a multilevel cache system utilized in conjunction with a multithread data processing system.